Medical image segmentation method based on self-supervised pre-training and two-stage fine tuning training

Through self-supervised pre-training and two-stage fine-tuning training, the accuracy of medical image segmentation is improved by using labelless data and public data, solving the problem of dependence on large-scale annotation data in the prior art, and achieving a more efficient medical image segmentation effect.

CN120431333APending Publication Date: 2025-08-05ARTIFICIAL INTELLIGENCE RES INST OF HEFEI COMPREHENSIVE NAT SCI CENT (ANHUI ARTIFICIAL INTELLIGENCE LAB)
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202510556781.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-29
Publication Date
2025-08-05

AI Technical Summary

Technical Problem

Existing medical image segmentation methods rely on large-scale precise annotation data, and fail to fully utilize the rich information in labelless data and public data, resulting in limited segmentation performance.

Method used

The self-supervised pre-training and two-stage fine-tuning training methods are adopted, and the preprocessed medical image data set, coarse-tuning public data set and fine-tuning target data set are used to build a segmentation model through a self-supervised pre-training network and U-Net decoder, and train it using a multi-stage loss function to improve segmentation accuracy.

Benefits of technology

In the scenario where medical image segmentation mask is insufficient, effectively using labelless data and public data is improved to improve the accuracy and performance of medical image segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120431333A_ABST
    Figure CN120431333A_ABST
Patent Text Reader

Abstract

The invention discloses a medical image segmentation method based on self-supervised pre-training and two-stage fine tuning training. The method comprises the following steps: 1, obtaining a pre-trained medical image data set, a coarse tuning public data set and a fine tuning target data set after preprocessing; 2, building a self-supervised pre-training model, training a self-supervised pre-training framework by using pre-training medical image data, and calculating loss of a reconstructed image and an original image by using a mean square error function to obtain a pre-training encoder; 3, constructing a segmentation model based on a pre-training encoder and a U-Net decoder, and performing supervision constraint on a prediction result by using a coarse tuning public data full-parameter coarse tuning segmentation model to obtain a coarse tuning model; and 4, performing supervision constraint on a prediction result by using a coarse adjustment model obtained by fine adjustment of all parameters of the fine adjustment target data, and predicting a medical image segmentation result. According to the method, the medical image segmentation performance is improved in the scene of insufficient medical image segmentation masks from the angles of self-supervised pre-training and two-stage fine adjustment.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Cited By

  • Large bridge apparent disease intelligent extraction method based on unmanned aerial vehicle inspection images

    CN121305380A

  • Multi-level land resource segmentation migration fine adjustment method based on input alignment

    CN121438099A

  • A multi-level land resource segmentation and migration fine-tuning method with input alignment

    CN121438099B